TiO₂-mediated visible-driven photocatalytic degradation of phenolic compounds: predictive modeling and optimization via machine learning techniques

Numerous machine learning (ML) models have been developed for optimizing and predicting the photodegradation of organic pollutants. However, their low predictive power restricts their applicability. In this work, three different ML models—random forests (RF), extremely randomized trees (ERT), and artificial neural networks (ANN)—were developed and evaluated in terms of predicting and optimizing the black TiO₂-mediated visible driven photodegradation of phenolic compounds (PHCs). The ERT model output displayed lower overall errors and higher R² indicating better performance in forecasting the photodegradation of PHCs. Using the ERT model, representative three-dimension (3D) plots produced the best visual representation of the estimates of the effects of different input variables on PHCs photodegradation. At optimal pH 7.3, 0.86 g/L of TiO₂, 0.039 mol/L of H₂O₂, and 33.45 mg/L of PHCs concentrations, the ERT model prediction was 73.35% photodegradation of PHCs, which is the closest to the actual mean (72.79%) suggesting the improved generalizability and predictive power of the model.